I Used AI To Write a 630+ Page Novel. Here’s What Actually Worked.
What started as a KDP experiment turned into 102,737 words, 32 chapters, an author site and the first book in a planned trilogy.
I recently got sucked into the KDP self-publishing world, and somewhere along the way I accidentally became a fiction author.
That wasn’t really the plan.
I’ve been in online marketing and digital products for more years than I care to count, so KDP was something I’d looked at before. But looking at a business model and producing a 630+ page fantasy novel are two very different things.
What began as an experiment turned into Paladin 404: Welcome to the Dungeon, a 102,737-word GameLit novel now published on Kindle.
Amazon US:
https://alink.co/paladin404us
Amazon UK:
https://alink.co/paladin404uk
Author site:
https://blogbum.com
It’s $0.99 as an early-bird launch price, going up to $2.99 after launch.
This post isn’t about persuading you to read a fantasy novel.
It’s about what happened behind the scenes.
I thought AI would make the writing the easy bit
You’ve seen the demonstrations.
Somebody types “write me a 50,000-word fantasy novel”, waits a few minutes, and announces they’ve written a book.
Technically, I suppose they have.
But I didn’t want 100,000 words of plausible-looking AI fiction.
I wanted characters who stayed consistent, a murder mystery where the clues actually made sense, and decisions made in Chapter 4 that were still true in Chapter 29.
The real challenge was never generating words. It was keeping control of a large creative project while using AI to accelerate parts of it.
That problem applies to a lot more than novels.
The idea itself was simple
Alf is older, not remotely heroic, and finds himself dumped into a fantasy dungeon where everyone else seems far more relaxed about it than he is.
He gets a homemade spear, a suspiciously useful notebook, and a mushroom creature called Fluffy who adopts him.
Everybody keeps insisting he’s some kind of Paladin.
He disagrees.
Fairly strongly.
What Alf is actually good at is noticing when things don’t add up.
So when creatures start appearing where they shouldn’t and somebody starts profiting from the chaos, he begins asking questions.
Then somebody is murdered.
Which gave me the line the book hangs on:
The Dungeon isn’t breaking its rules. Someone is using them.
I didn’t need AI to invent fantasy events.
I needed a system that could build and protect a story.
I stopped treating AI like an author
That was probably the most important decision in the entire project.
Instead of asking AI to “write my novel”, I started treating it like a production team.
– Story Bible
– World Bible
– Character continuity
– The rules of the Dungeon and its System
– The mystery structure itself
We tracked what every character knew, what they didn’t, and which clues had landed.
Even the comedy had continuity files.
The word that mattered most was lock.
Once I approved something, it became manuscript truth.
AI could offer alternatives before that point. Afterwards, everything downstream had to respect the decision unless I explicitly changed it.
That single rule stopped the project endlessly rewriting itself.
The surprise was that my detailed plan lost.
Once scenes were actually written, some characters turned out far more interesting than they’d been on paper.
So the outline became a navigation system rather than an instruction manual, and the planning files were updated to match the book.
Not the other way around.
Without that rule, an AI-assisted project gets trapped inside its own paperwork.
Thirty-two chapters in, I found I still hadn’t finished the book
The first complete draft came in just under 96,000 words.
Book finished.
Except it wasn’t.
Instead of formatting it and throwing it at Amazon, I went back through the whole thing in eight revision batches of four chapters each, then audited the entire manuscript again as a single object.
That final pass wasn’t a rewrite.
It was closer to regression testing software.
Had a character’s voice drifted?
Was somebody carrying an object they’d left behind three chapters earlier?
Did every clue still appear before the deduction that used it?
It also caught three places where the narration had accidentally referred to the book as a book.
Those are the ones that make you go quiet for a minute.
The manuscript finished at 102,737 words, and every extra word had to earn its place.
Finishing the story and finishing the book are not the same job.
Generating 100,000 words is no longer difficult.
Turning them into something coherent is where the work starts.
Then I had to sell the right book
I’d been using LitRPG and GameLit interchangeably.
But the finished book doesn’t have the heavy stat screens many LitRPG readers expect.
Calling it LitRPG might have won more clicks.
It would also have attracted the wrong reader.
So the positioning settled on two phrases:
A Dungeon Detective GameLit
and
Monsters, Murder & a Mushroom
One tells you what shelf it belongs on.
The other tells you why this one is worth noticing.
I’ve spent decades in direct response, so optimising for the biggest promise would have been easy.
But if your marketing attracts somebody who wanted a different book, all you’ve done is accelerate the bad review.
Package the actual reading experience, not the market opportunity you wish you had.
The cover turned into a project of its own
We tested genuinely different directions rather than twenty versions of “man standing in a dungeon”.
AI made variation ridiculously cheap.
It also kept trying to help.
It invented details that weren’t in the book, added jokes the story hadn’t earned, kept redesigning Fluffy, and polished Alf’s deliberately terrible homemade spear into a proper fantasy weapon.
Sometimes it produced something that looked great and sold a completely different story.
Plausible is not the same as correct.
The test that mattered in the end was brutally simple.
Shrink the cover to the size somebody actually sees while scrolling Amazon.
Can you still tell what it is?
Nobody buys your beautiful 2560-pixel artwork.
They buy the tiny rectangle they noticed next to forty other tiny rectangles.
Amazon sells the book. The author site builds the audience.
The book needed an author, and I published as blogbum — partly because it suits the tone, mostly because I already owned the domain.
Which meant the pen name could be an asset rather than a line on a listing.
So I used AI to build https://blogbum.com as the home of the series.
Amazon is where the transaction happens.
The author site is where the relationship starts.
The free short stories live there, readers can meet the world without buying anything first, and I get to build a direct audience of people who might want Book Two.
That thinking also produced https://ShortStoryEngine.com — because writing a book and then desperately hunting for readers is backwards.
So a book-writing experiment has produced a novel, an author brand, an owned site, free supporting fiction and a separate software idea.
None of which was on the original plan.
So did AI write my book?
AI was involved at nearly every stage — ideas, world, project files, prose, continuity and cover concepts.
But “AI wrote it” is a hopelessly inadequate description of what happened.
I decided what the book was, what became canon, which prose was approved, which jokes survived, what the cover promised, and when the book was finished.
AI is extremely good at accelerating production. Acceleration is only useful when you still control the destination.
I’ve documented the whole thing properly
I got carried away documenting this and ended up with a full Paladin 404 case study PDF — the framework I adapted, the continuity systems, the chapter-production workflow, the eight-batch revision programme, the whole-book audit, the cover development, the KDP packaging, the mistakes, and every place AI produced something perfectly plausible and completely wrong.
I’ve also recorded a video walkthrough showing how I used AI to build the blogbum.com author site.
I’m not putting either out publicly.
They’re a bonus for people who buy the book.
- Grab the book:
US: https://alink.co/paladin404us
UK: https://alink.co/paladin404uk - Send me your receipt or purchase details:
Facebook: https://facebook.com/matgarrett
Support: https://gazmat.com - I’ll send you the case study PDF and the author-site build video.
That way you get both sides of the experiment.
The finished product, and the system behind it.
The next part of the case study is the numbers
There’s one section I can’t write yet.
The results.
Sales, page reads, reviews, site traffic, short-story readership, email subscribers, and eventually whether building an audience around Book One makes launching Book Two any easier.
That’s the question I’m most interested in:
Can you build demand for a series before the next book exists?
Now I get to find out.
Paladin 404: Welcome to the Dungeon is on Kindle now:
Amazon US: https://alink.co/paladin404us
Amazon UK: https://alink.co/paladin404uk
You can also read the free short stories and see what I’m building around the series at:
And if you actually read the book rather than merely buying it for the case study, tell me what you make of Alf.
After this long tracking what he, Mavis and a small mushroom know at any given moment, human feedback is the most useful thing I can get.
